User Guide

Built-in Robots & Datasets

Robot descriptions and datasets live in the HuggingFace dataset openrobokit/robokit.

Assets are under their own licenses; see license files and outstanding gaps.

Consuming

Import a named asset module and use its static Path symbols. Required files are downloaded and cached automatically. Use fetch() for paths without static symbols:

python

from robokit.assets import fetch
from robokit.assets.robots.humanoids import unitree_g1

unitree_g1.URDF_PATH
unitree_g1.COLLISION_SPHERE_PATH
fetch(["datasets/sfu/**"]) / "datasets/sfu/0005/x.npz"

from robokit.assets import fetch
from robokit.assets.robots.humanoids import unitree_g1

unitree_g1.URDF_PATH
unitree_g1.COLLISION_SPHERE_PATH
fetch(["datasets/sfu/**"]) / "datasets/sfu/0005/x.npz"
assert unitree_g1.URDF_PATH.is_file()

Note

Files are pinned to one dataset commit and cached under ~/.robokit/, so imports work offline.

Self-collision ignore sets

Each robot lists the link pairs that always touch, so self-collision terms skip them. Each geometry representation has its own set. Match it to your SelfCollisionTask representation (sphere by default):

python

robot = Robot.load(
    xhand.URDF_PATH_RIGHT,
    load_collision_spheres=True,
    collision_spheres_path=xhand.COLLISION_SPHERE_PATH_RIGHT,
    self_collision_ignore_path=xhand.SPHERE_SELF_COLLISION_IGNORE_PATH_RIGHT,  # or CAPSULE_...
)

from robokit.robo import Robot
from robokit.assets.robots.end_effectors import xhand

robot = Robot.load(
    xhand.URDF_PATH_RIGHT,
    load_collision_spheres=True,
    collision_spheres_path=xhand.COLLISION_SPHERE_PATH_RIGHT,
    self_collision_ignore_path=xhand.SPHERE_SELF_COLLISION_IGNORE_PATH_RIGHT,  # or CAPSULE_...
)
assert robot.spec.has_collision_spheres

MJCF variants have only MJ_CAPSULE_... sets. Robots without collision spheres have no sphere set.

For asset layout and authoring, see the Development Guide.

Robots

Licenses are upstream. — means unavailable or unspecified. A side, such as (right), means only that side is available.

Arms

Name	Format	Source	License	Collision spheres	Self-collision ignore
`franka_panda`	URDF	ManiSkill	Apache-2.0	YAML (NVIDIA non-commercial)	✓
`franka_panda`	MJCF	Menagerie	Apache-2.0	—	✓
`xarm6`	URDF	UFACTORY	BSD-3-Clause	YAML	✓
`xarm7`	URDF	UFACTORY / SAPIEN	BSD-3-Clause + —	—	✓
`xarm7`	MJCF	Menagerie	BSD-3-Clause	—	✓
`yumi`	YAML	—	—	YAML	✓

End effectors

Name	Format	Source	License	Collision spheres	Self-collision ignore
`ability_hand`	URDF	PSYONIC	MIT	YAML (right)	✓ (right)
`allegro_hand`	URDF	SimLab	BSD-style	YAML (right)	✓ (right)
`arx_x5_gripper`	URDF	RoboDojo	MIT	YAML	✓
`barrett_hand`	URDF	Barrett Technology	BSD-style	YAML	✓
`dclaw_gripper`	URDF	ROBEL	Apache-2.0	YAML	✓
`inspire_hand`	URDF	Inspire Robots	CC-BY-NC-SA-4.0	YAML	✓
`leap_hand`	URDF	Ananye Agarwal	MIT	YAML (right)	✓ (right)
`mano_hand`	URDF (right)	MANO	MANO	YAML (right)	✓ (right)
`orcahand`	URDF (right)	ORCA	MIT	YAML (right)	✓ (right)
`paxini_hand`	URDF	PaXini	—	YAML (right)	✓ (right)
`robotiq_2f85`	URDF	a-price	BSD	YAML	✓
`schunk_hand`	URDF	SCHUNK	GPLv3	YAML (right)	✓ (right)
`shadow_hand`	URDF	Shadow Robot	BSD-3-Clause + Apache-2.0	YAML (right)	✓ (right)
`shadow_hand`	MJCF (right)	Menagerie / DexGraspNet	Apache-2.0 + CC-BY-NC-4.0	YAML (right)	—
`sharpa_hand`	URDF	Sharpa Robotics	Apache-2.0	YAML	✓
`umi_gripper`	MJCF	Menagerie	MIT	—	—
`xhand`	URDF	RobotEra	—	YAML (right)	✓ (right)
`xhand`	MJCF (right)	RobotEra	—	—	✓
`wuji_hand` (Hand 1)	URDF	Wuji Technology	MIT	YAML	✓
`wuji_hand` (Hand 1)	MJCF	Wuji Technology	MIT	—	—
`wuji_hand_2_2` (Hand 2.2)	URDF	Wuji Technology	MIT	YAML	✓
`wuji_hand_2_2` (Hand 2.2)	MJCF	Wuji Technology	MIT	—	—
`yam_gripper`	URDF	I2RT	MIT	YAML	✓
`yam_gripper`	MJCF	I2RT	MIT	—	✓

Humanoids

Name	Format	Source	License	Collision spheres	Self-collision ignore
`berkeley_humanoid_lite`	URDF	Hybrid Robotics	CC-BY-SA-4.0	YAML	✓
`booster_k1`	URDF	Booster	BSD-3-Clause	YAML	✓
`booster_t1`	URDF	Booster	BSD-3-Clause	YAML	✓
`engineai_pm01`	URDF	EngineAI	BSD-3-Clause	YAML	✓
`fourier_gr3`	URDF	Fourier	GPLv3	YAML	✓
`pal_talos`	URDF	PAL Robotics / Stack of Tasks	LGPL-3.0	YAML	✓
`pnd_adam_lite`	URDF	PNDbotics	BSD-3-Clause	YAML	✓
`tienkung_lite`	URDF	X-Humanoid	OpenAtom Open Hardware 1.0	YAML	✓
`toddlerbot_2xc`	URDF	ToddlerBot	MIT	YAML	✓
`unitree_g1`	URDF	Unitree	BSD-3-Clause	YAML	✓
`unitree_h1`	URDF	Unitree	BSD-3-Clause	YAML	✓
`unitree_h1_2`	URDF	Unitree	BSD-3-Clause	YAML	✓

Datasets

Name	Format	Source	License
`climb`	NPY, OBJ	OmniRetarget / Holosoma	Apache-2.0
`dex_retargeting`	PKL	dex-retargeting / DexYCB	CC-BY-NC-4.0
`dexmachina`	NPZ	DexMachina / ARCTIC	ARCTIC
`dexycb`	NPZ, JSON, OBJ	DexYCB	CC-BY-NC-4.0
`gigahands`	NPZ	GigaHands	CC-BY-NC-4.0
`nist_assembly`	XML, NPZ	AssembleNIST / UW Lab	MIT
`oakink2`	NPZ	OakInk-v2	CC-BY-SA-4.0
`omomo`	PT, PKL, OBJ	OMOMO / InterMimic	—
`parahome`	PKL, JSON, OBJ	ParaHome	CC-BY-NC-SA-4.0

User-provided body models

Download MANO or SMPL-X from its official website after accepting its license. Put the files you need under ~/.robokit/body_models/, or set ROBOKIT_BODY_MODELS to another folder:

mano/MANO_RIGHT.pkl or mano/MANO_LEFT.pkl
smplx/SMPLX_NEUTRAL.npz, smplx/SMPLX_MALE.npz, or smplx/SMPLX_FEMALE.npz

python

from robokit.smplx import load_mano, load_smplx

hand = load_mano()  # or load_mano("/path/to/MANO_RIGHT.pkl")
body = load_smplx()

from robokit.smplx import load_mano, load_smplx

hand = load_mano()  # or load_mano("/path/to/MANO_RIGHT.pkl")
body = load_smplx()
assert hand is not None and body is not None

NPZs load directly. PKLs convert once beside the originals, so the folder must be writable. RoboKit never downloads body models.

← PreviousHand-Eye Calibration Next →Robot